519880

research-article2014

SOC0010.1177/0038038513519880SociologyMills

Class Debate

The Great British Class Fiasco:

A Comment on Savage etal.

Sociology
2014, Vol. 48(3) 437444
The Author(s) 2014
Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0038038513519880
soc.sagepub.com

Colin Mills

University of Oxford, UK

Abstract
Savage etal. (2013) claim they have produced a new model of the British class structure. They
stress the innovative use of an internet survey, the BBCs Great British Class Survey (GBCS), but
it plays no serious role in the generation of their class typology. What they do is a theory free
(though Bourdieu inspired) data dredging exercise. What they derive is an arbitrary typology
determined by a contingent fact the size of their sample.

Keywords
Bourdieu, social class, stratification

Introduction
This comment criticizes Savage etal. (2013) published in this journal. My argument is in
five sections. Firstly, I outline the standard view of typologies and compare it to Savage
etal.s. Secondly, I deal with their data collection strategy and in particular the limit it
imposes on their ability to make statistical inferences. Thirdly, I point out the implications of their model selection strategy. Fourthly, I reveal their inability to interpret their
own results correctly and raise some additional questions about the operationalization of
their concepts. Finally, I discuss the substance of their typology and draw out some of the
bizarre consequences that the authors must accept if they wish to claim consistency for
their model.

On Typologies
Concepts are abstract entities. We use them to selectively organize experience. Typologies
are closely related to concepts; they are the step we take in order to make abstract ideas
Corresponding author:
Colin Mills, Nuffield College, University of Oxford, Oxford OX1 1NF, UK.
Email: colin.mills@sociology.ox.ac.uk

Downloaded from soc.sagepub.com by guest on January 13, 2015

438

Sociology 48(3)

more concrete. There can be as many typologies as there are points of view. We are free
to fill a concept with whatever content we like and then derive a typology from it.
Typologies should be made for a purpose and we should ask both whether that
purpose is worthwhile and whether the proposed typology is fit for purpose. Though
we are free to organize what we can learn of other peoples circumstances with any
conceptual scheme we like, there are logical constraints, and we must acknowledge
what follows from doing things one way rather than another. I return to this point in
Section 6.
Some conceptual schemes are more useful than others; they, as Plato put it, cleave
nature at its joints. Normally the purpose is given by the part that a concept (and its associated typology) plays in a theory that explains or predicts something and naturally this
entails spelling out what it is that the concept is to play a part in explaining or predicting.
From this it follows that a typology is a practical tool. Inventing a typology for its own
sake is pointless.
Let me give two examples. In market research, market segmentation is a procedure
that builds predictive models to help clients target products at people that are likely to
buy them. In this kind of exercise all that matters is that good predictions are produced;
the prediction model itself can be built in an entirely inductive manner. The categories
that are produced, which may be an ad hoc combination of demographic, lifestyle and
dispositional characteristics, are validated by their predictive value.
Now consider John Goldthorpes most recent statement of what social class is about
(2000: 20629), a view that underpins the construction of the National Statistics Socioeconomic Classification (NS-SEC) (Rose and Pevalin, 2003). It focuses on selected
aspects of the employment relationship and ignores others.
I have played a small part in empirically examining the correspondence between these
things and the occupational categories that form part of the NS-SEC. In McGovern etal.
(2007), together with my colleagues, I argue that in principle there is no reason why
social class indicators should not be measured at the level of the individual job and that,
if they were, the categories of the NS-SEC would be redundant.
In practice, this is unrealistic. No general purpose survey will collect the detailed
information required for accurate measurement of the real variables of interest and
thus it makes sense to chunk the world of employment up into occupations and
then aggregate these into classes as long as we have evidence that these classes
serve as adequate proxies for the underlying variables that practicalities dictate we
cant measure directly. Goldthorpe would, no doubt, logically justify the classification of the underlying dimensions by saying that he is creating a tool that can be
used in situations and with data where there are no direct measures of the things he
is interested in. By well-validated groupings of occupations he can capture a substantial part of what he means by social class. Both the market researchers and
Goldthorpe can justify their classifications by their objectives. Lets now ask
whether Savage etal. can do the same. What do they expect their class typology to
explain?
A close reading of the article itself gives no answer. The closest I can come to a clue
is a statement in the Guardians Comment is free section by the lead author:

Downloaded from soc.sagepub.com by guest on January 13, 2015

439

Mills

The concept of class matters because we need a way of connecting accentuating economic
inequalities to social and cultural differences which permeate our society. Rather than seeing
our lifestyles and social networks as somehow separate from economic inequalities, there are
overlaps that can work together to produce social advantage and disadvantage. (Savage, 2013)

That lifestyles, features of social networks, cultural differences and economic inequalities are correlated is not something that many would doubt but the demonstration of such
a banal fact surely cannot be the principal objective of Savage etal.s efforts. What is
clear is that their typology cannot have any role in explaining differences in the distribution of social and cultural capital for the simple reason that all of these things are built
into the definition of their typology.

On Sample Selection
The GBCS is a self-completion questionnaire administered over the internet. Internet
surveys are known to be subject to self-selection bias. Savage etal. acknowledge this and
demonstrate its consequences for the GBCS by benchmarking against the NS-SEC typology they elsewhere affect to disdain. Managers and professionals are vastly overrepresented. Applying post-stratification weights is not a viable option because this only
corrects for selection on variables with a known (or reliably estimable) distribution in the
population of interest. Some variables of relevance primarily the standard demographic
face sheet variables are available, others the distribution of cultural and social capital
are not, yet are obviously mission critical. Thus an additional source of information
about the observable sources of response bias is needed. Enter the GfK survey.
Despite having over 160,000 responses to the GBCS, the information that Savage
etal. base their empirical claims on is a quota sample (the so-called GfK sample) of 1026
respondents. This is a flimsy source upon which to build their inductive edifice but without it the GBCS data is essentially useless. In fact it turns out that even with it the GBCS
is less valuable than they lead us to believe.
Given that the quota sample is the bedrock of the whole enterprise, and the GBCS is
for analytical purposes set aside, we should be told something substantial about it. All we
are in fact told is that that both GfK and the authors are satisfied that the demographics
are nationally representative.
I could go on to outline the well-known problems of making inferences from sample
data selected by a non-probability method but I will leave this for others to take up.
Suffice to say that with a probability sample there is a coherent procedure for making
sample to population inferences, including, most importantly, inferences about uncertainty: with a quota sample there is none, unless one is prepared to assert that the correct
data generating model has been identified. Savage etal. have a mountain of highly selfselected poor quality data sitting on top of a molehill of (slightly) better quality data.
Ultimately they want to use the latter to make a sensible calibration of the former, but
they have to do this without any basis for assessing the uncertainty surrounding the numbers they estimate.
To take just one example of the difficulties this creates, much fuss is made about the
discovery of a social class category they call the elite. They tell us that this is about 6

Downloaded from soc.sagepub.com by guest on January 13, 2015

440

Sociology 48(3)

per cent of the population. If the data were from a simple random sample a rough 95%
confidence interval would be 1.5 per cent. But that is foolishly optimistic and takes no
account of any other source of variance inflating error. One could easily double this
number. So the elite could, on the basis of these data, be 3 per cent of the population,
but it could also be 9 per cent of the population. Nine per cent is so large that Savage
etal. might want to think again about the appropriation of the elite label. They have no
basis, other than wishful thinking, for favouring either number (or anything in between).
The seven social classes that Savage etal. find fall out of a so-called latent profile
model estimated on the GfK quota sample plus one case consisting of the 161,400
responses to the GBCS each given a weight of 1/161400. This means that the parameter
estimates that determine how respondents are allocated to social classes are derived
almost entirely from the quota sample. The GBCS respondents are then allocated to the
latent class categories on the basis of the model. There will be a considerable amount of
uncertainty in doing so because the parameters of the latent profile model will not be
well determined, but lets suspend our disbelief for there is another snag.
Even if we accept that the model enables us to allocate individuals in the GBCS data
to the right classes, it does not allow the relationship between the class categories and
any variable in the GBCS external to the latent profile model to be estimated without
bias. This is important, for it is in fact what Savage etal. go on to do; for instance, when
they tabulate data on occupations conditional on class membership, or when they map
the density of class membership by geographical region. Calibration of the GBCS by the
GfK only works for variables that have been observed in both surveys but it cannot make
an appropriate adjustment for selection into the GBCS on the basis of unobservables.
Imagine a population in which we are able to calculate scores on a latent variable by
combining a suitable statistical model with information on three observed variables, say
economic capital, cultural capital and social capital. Lets call this variable X. We also
have information on two other variables. The first combines all the factors that influence
the propensity of an individual to respond to internet-based surveys. For the sake of concreteness lets call it curiosity or C for short. The second is a variable of interest that is
not included in our latent variable model; lets call it Y, for example a simple 1/0 dichotomy indicating whether or not an individual is a university graduate.
In the population it is straightforward to calculate P(Y=1|X), the probability that Y
equals 1 conditional on X. Now assume the following simple selection rule: there is a
threshold T such that if C>T an individual will respond to an internet survey and if C<
T an
individual will not respond. Further assume that X, C and T are all positively correlated.
When we carry out an internet survey what we observe is P(Y=1|X,C>T), the probability that Y equals 1 conditional on X and that C is greater than the threshold value and
this will not, except in uninteresting cases, be equal to P(Y=1|X). The size of the difference will depend on the strength of the inter-correlation between X, C and Y and the
point at which we set T.
We know that there is strong self-selection into the GBCS on the basis of observable
characteristics and it is implausible to assume that there is no self-selection on the basis
of unobserved curiosity. Even if we correctly assign a GBCS respondent on the basis
of the GfK to be a member of the traditional working class they would, on average,
have a curiosity score that was higher than the average in their class and because C is

Downloaded from soc.sagepub.com by guest on January 13, 2015

441

Mills

correlated with Y they would also have a higher than average propensity to be a graduate
than a non-graduate.
The magnitude of the bias induced by selection into the GBCS on the basis of unobservables is unknown and, I believe, not estimable from these data. It would, however,
be heroic to assume it didnt exist and doubly heroic to assume that it wasnt compounded by a similar selection bias affecting inclusion in the GfK quota sample.
Many readers of this journal might not appreciate the implications of the preceding
argument. To put it in a nutshell, it is highly likely that any comparison between Savage
etal.s class categories with respect to variables contained in the GBCS but not used to
derive the class categories such as educational attainment, type of university attended,
geographical location and so forth will be biased to a degree that may be substantively
significant.

On Model Selection
We are told little by Savage etal. about the discovery of their seven social classes.
They state that the Bayesian Information Coefficient is minimized when seven latent
classes are assumed. We are not told how much worse things are when six or eight
latent classes are assumed, or the extent to which these solutions resemble the seven
class solution. We are given no sense whatsoever of model uncertainty. In fact there is
a dark secret. The selection of seven classes is a function of the sample size of the GfK
survey and without well-articulated theoretical grounds for distinguishing the classes
it could not be otherwise. It is just as well that Savage etal. discard the GBCS data. If
they had estimated their latent profile model with 161,400 cases they would conclude
that there were rather more than seven classes. The number of classes is a consequence
of a decision about how much data to collect. This does not look like cleaving nature
at its joints.
Savage etal. will no doubt claim that they are pursuing an open-minded inductive
strategy but this simply wont do. A sensible model selection strategy must be based not
only on formal statistical criteria but also on whether the model makes sense in terms of
the configuration of variables that go into it; but it is just too facile to pass off unconstrained post-hoc interpretations as though they are facts of nature. This is particularly
true when one allows oneself, as Savage etal. do, to go on a fishing expedition that
encompasses variables external to the latent profile model itself. It is permissible to use
external variables to validate a typology once you have made it, but peeking at them
while making decisions about whether there are six, seven, eight or 77 latent classes is
having your cake and eating it.

On Measurement
In the course of discussing Goldthorpes class schema Savage etal. say:
the schema has been shown to be of less use in explicating the wider cultural and social
activities and identities which do not appear to be closely linked to peoples class position,
as defined by the Goldthorpe class schema (2013: 222)

Downloaded from soc.sagepub.com by guest on January 13, 2015

442

Sociology 48(3)

This is a bold claim, but it is not supported by Savage etal.s own Figure 1 (2013: 227).
Following a line from NW to SE is a dimension that puts high-brow cultural taste at the
NW pole (people visit stately homes, go to the theatre, museums and galleries, etc.) and
low-brow taste in the SE corner. Their Figure 2 (2013: 228), which projects supplementary points into the space of Figure 1, shows that the categories of the conventional
NS-SECs fall from top to bottom roughly in a NW to SE fashion. This means that they
are indeed predictive of the first cultural dimension.
Savage etal.s initial belief is wrong. We are told that one of the motivations for the
project is that conventional class categories are inadequate for the sociological analysis
of the sorts of cultural tastes and practices they are interested in. But when their own data
show that they are mistaken, rather than abandoning their initial premise they carry on
regardless.
What of the second dimension in Figure 1? This is given the label emerging cultural
capital though in what sense it is emerging and what it is emerging from we are never
told. It runs from SW to NE and it distinguishes the sorts of things that young people like
doing from the sorts of things that older people dislike or dont do. There is no doubt that
younger people do and enjoy things that older people dont but what has this to do with
social class? Do peoples tastes and activities change as they get older? Yes. Do peoples
tastes depend on when they were born? Yes. What dimension 2 represents is a mixture of
life-cycle and cohort differences. Nobody denies that the young and old are different. Its
a novel claim, however, to assert that they are different social classes. I return to the
implications of this absurdity in section 6.
Savage etal. are to be commended for acknowledging that wealth and not just income
is important to their concerns, but they underestimate the difficulty of collecting reliable
wealth data. Respondents to the GBCS are asked to provide information about savings
and pensions. However, expecting respondents to provide an accurate estimate of the
value of their pension is a stretch too far.
Before we even consider whether their assets are actually liquid (try as a 30-year-old
to borrow money against future pension income), we have to confront the difficulty
respondents will have in calculating the assets present discounted value. This will
involve choosing a suitable discount rate and making a guess about how long they are
going to survive after reaching pensionable age. In principle there is no difficulty in
arriving at a sensible number: in practice it strains credibility to assume that many people
will have bothered to do it.

On Social Classes
In this section I invite a further suspension of disbelief. Lets pretend that there are convincing replies to all of the points Ive made so far and we take Savage etal.s seven
classes seriously.
Difficulties immediately arise when Savage etal. interpret the seven social classes
they discover (Tables 5 and 6, p. 230). Normally one would do this in terms of the profiles of the observed variables associated with each of the latent classes. But looking at
Table 6, it becomes apparent that this is not straightforward. There are differences
between the classes but which differences are important?

Downloaded from soc.sagepub.com by guest on January 13, 2015

443

Mills

In fact the interpretation and certainly the labels applied to the classes actually come
from information quite extraneous to the latent profile analysis itself (Table 7, p. 231).
Bizarrely, one of the key ingredients is the percentage classified in the NS-SEC categories that Savage etal. claim are inadequate.
Age also seems to play a prominent part in the construction of the classes. While this
would be unobjectionable if the objective was market segmentation, it is difficult to see
how it can be reconciled with conventional understandings of social class. Usually,
classes are thought of as groups within which individuals could potentially spend the
whole of their lives. However, Savage etal.s inductive method allows classes to be distinguished by tastes and activities that have a strong age gradient. The implication is that
individuals can either grow out of their class because they become too old to get down to
the gym, or become stuck because of the cultural tastes they acquired during early
adulthood.
Table 7 is a critics gold-mine. Take for instance the traditional working class. The
average age of this group is 66! They have modest incomes, are low on emerging cultural capital, yet 30 per cent have, or had, jobs that would be classified as professional
or managerial. A large proportion of this group are pensioners. Pensioners are an important interest group, but what analytical insight do we get from calling them a social class?
Life cycle plays a role in distinguishing what Savage etal. term the elite and the
established middle class. The latter differ from the former principally by having smaller
incomes, less savings and living in cheaper houses. In terms of social and cultural capital
they dont differ at all. When we look at Table 7 we find that the established middle
class are on average 11 years younger than the elite and other differences are quite marginal. A substantial part of the difference could plausibly be attributed to life-cycle stage
and having a stake in the south-easts housing market. Again we must ask: what is gained
by relabeling age and now geography as class?
Life cycle also plays a role in the identification of the emergent service workers.
These are young people who go to gigs, join a gym and watch sport. They have few savings and little invested in home ownership. If you were to say that this is a lifestyle
group, say young single people or young childless couples then there would be little to
object to: but a social class? Its a strange sort of social class that people will grow out of
simply by ageing, getting married and having kids.
In short, Savage etal. fail to take the content of their typology seriously. If they did
they would have to confront consequences which should give them cause to rethink their
approach. For example, they would have to accept that people could change their social
class at will simply by turning off the Beatles and turning on to Beethoven. Of course
people do change class but if it was as easy as this the solution to Britains so-called
social mobility problem would have been spotted long ago.

Conclusions
My conclusion is that for the reasons I outline here, and for others that space limitations
prevent me from mentioning, the GBCS is a fiasco. It is so theoretically and methodologically flawed that it can contribute little of value to our understanding of the structure
of systematic social inequality in the UK.

Downloaded from soc.sagepub.com by guest on January 13, 2015

444

Sociology 48(3)

I want to make it easy for Savage etal. to reply to my criticisms so I will end with the
questions I think they need to address: What is your typology meant to explain? Why
should we have confidence in a typology built on the basis of such a small amount of
data? What will you do when your method is applied to a larger amount of data and you
discover, as you undoubtedly will, a larger number of classes? What use can the GBCS
(as opposed to the GfK) data actually be put to? Do you accept that your data show that
cultural consumption is related to conventional measures of social class? What is gained
by relabeling age-cohort and life-cycle stage as social class? And finally, do you really
believe that changing ones social class can be a matter of getting out of bed and making
a serious effort to like Brahms or to attract a few more Facebook friends?
Acknowledgements
I would like to thank Robert Erikson, Geoff Evans, John Goldthorpe, Eric Harrison, Jouni Kuha,
Kenneth Macdonald, Patrick McGovern and David Rose for their extremely helpful comments and
advice about which of many lines of argument to pursue. Unfortunately the word limits imposed
on a comment mean that many of these, including some I had promised to develop, have had, in
the event, to be dropped.

Funding
The writing of this comment received no specific grant from any funding agency in the public,
commercial or not-for-profit sectors.

References
Goldthorpe JH (2000) On Sociology. Oxford: Oxford University Press.
McGovern P, Hill S, Mills C and White M (2007) Market Class and Employment. Oxford: Oxford
University Press.
Rose D and Pevalin DJ (eds) (2003) A Researchers Guide to the National Statistics Socioeconomic Classification. London: Sage.
Savage M (2013) Concerned about the BBCs Class Calculator? Let me explain. The Guardian,
Comment is free, 10 April. Available at: http://www.guardian.co.uk/commentisfree/2013/
apr/10/bbc-class-calculator
Savage M, Devine F, Cunningham N, etal. (2013) A new model of social class? Findings from the
BBCs Great Class Survey experiment. Sociology 47(2): 21950.
Colin Mills is a Reader in the Department of Sociology, University of Oxford and a Professorial
Fellow of Nuffield College.
Date submitted September 2013
Date accepted November 2013

Downloaded from soc.sagepub.com by guest on January 13, 2015